Skip to main content

Cap AI spend for a Slack/support bot in 10 minutes

Talon gives EMEA SMB operators a simple promise:

You can cap AI spend before it runs away, and prove every allow/deny decision with signed evidence.

This guide shows the fastest path to that outcome for a support bot caller.

What this guide covers​

  • Hard daily/monthly spend caps for one caller (support-slack-bot)
  • Deny before any upstream provider call when the next request would exceed cap
  • Signed evidence for both allowed and denied decisions
  • Dashboard + CLI visibility for today/month, by caller/model/provider

1. Add a caller with hard EUR caps​

Use a gateway config like this (from examples/gateway/talon.config.gateway.yaml):

gateway:
  enabled: true
  listen_prefix: "/v1/proxy"
  mode: "enforce"

  callers:
    - name: "support-slack-bot"
      tenant_key: "talon-gw-support-xyz"
      tenant_id: "default"
      policy_overrides:
        max_daily_cost: 10.00
        max_monthly_cost: 200.00
        pii_action: "warn"
        allowed_models: ["gpt-4o-mini"]

For a demo, temporarily lower max_daily_cost (for example 0.01) so you can trigger a denial quickly.

2. Start Talon gateway​

talon serve --gateway --gateway-config=path/to/talon.config.yaml

Ensure provider keys are in vault (for example talon secrets set openai-api-key "sk-...").

3. Run the 6-step demo flow​

  1. Configure very low max_daily_cost for support-slack-bot
  2. Send one request that is allowed
  3. Send a second request that would exceed budget
  4. Verify Talon denies before provider call
  5. Verify signed evidence exists for the denial
  6. Verify dashboard budget utilization reflects the event

Example calls:

# Allowed request (first call)
curl -sS "http://localhost:8080/v1/proxy/openai/v1/chat/completions" \
  -H "Authorization: Bearer talon-gw-support-xyz" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-4o-mini","messages":[{"role":"user","content":"Summarize ticket #123"}]}'

# Denied request (second call, over cap)
curl -sS "http://localhost:8080/v1/proxy/openai/v1/chat/completions" \
  -H "Authorization: Bearer talon-gw-support-xyz" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-4o-mini","messages":[{"role":"user","content":"Summarize ticket #124"}]}'

Expected deny characteristics:

  • HTTP 403
  • Machine-readable error code/reason contains budget_exceeded
  • No upstream provider execution for denied call
  • Signed evidence row recorded with denial reason and estimated pre-call cost

4. Verify with CLI and dashboard​

CLI:

talon costs --tenant default
talon costs --caller support-slack-bot --json
talon costs --by-provider --tenant default

Export cost rows (joinable to signed evidence by evidence_id):

talon costs export --tenant default --caller support-slack-bot --format csv
talon audit export --tenant default --caller support-slack-bot --format signed-json

HTTP/API equivalents:

  • GET /v1/costs
  • GET /v1/costs/budget
  • POST /v1/costs/export

Dashboard:

  • /gateway/dashboard for real-time caller/model/provider operational view
  • /dashboard for evidence-backed governance view and drill-down

Cost visibility vs caps vs evidence attribution​

  • Cost visibility: today/month totals and breakdowns by tenant/caller/model/provider
  • Hard budget caps: deny requests that would exceed daily/monthly cap before provider call
  • Evidence-backed attribution: every allow/deny is signed and traceable by evidence ID, including budget-denied rows with zero provider cost and denial reason

This is the launch narrative connection to evidence integrity:

Verified evidence proves both governance decisions and cost attribution.

Native agents (no gateway)​

If you run talon run directly, use .talon.yaml:

policies:
  cost_limits:
    per_request: 0.50
    daily: 20.00
    monthly: 400.00

Next steps​

I want to…Doc
Verify signed exports and tamper checksHow to export evidence for auditors
Understand dashboard metrics schemaGateway dashboard
Add more callers/modelsConfiguration and environment
Apply additional governance snippetsPolicy cookbook